Embodiment Breaks the Tie: Identifying Whether the Model or the Grounder Failed in Neuro-Symbolic Planning

Embodied neuro-symbolic planners pair a symbolic action model, increasingly elicited from a large language model (LLM), with a learned grounder. When prediction and perception disagree, which component failed? We show that a systematic grounding error, such as a reasoning shortcut, and an action-model error produce identical execution logs, so repairing the model to fit passive traces damages correct clauses whenever the grounder is at fault. The tie can be broken by acting. Our method, PWA-AT, casts attribution and repair as one weighted MaxSAT problem over calibrated confidence and clause provenance; on a near-tie it takes the most informative discriminating action per unit cost, and an LLM proposes repairs that the solver validates against the trace and LTLf safety constraints. On GRIT, a 3D-scene-graph benchmark with attribution labels known by construction, with a correct model and injected systematic errors of the kinds PWA-AT searches over, PWA-AT succeeds in 0.872 of episodes (0.215 above the strongest baseline) and damages 0.011 of correct clauses, against 0.042--0.057 for the editing baselines; without acting, it damages more than they do.